BuildsGPUsModelsGuidesMethodologyNewsletter

What GPU do you need to run gpt-oss-20b?

20.91B parameters, 3.6B active (mixture-of-experts) (Hugging Face, accessed 2026-10-15). Q4_K_M GGUF: 11.67 GB (source).

Single GPU that fitsVRAMSpeed estimatePrice
NVIDIA GeForce RTX 508016 GB~363 tok/slow $1,600, avg $2,399 (new) source, 2026-10-16
NVIDIA GeForce RTX 5070 Ti16 GB~339 tok/slow $1,120, avg $1,792 (new) source, 2026-10-16
NVIDIA GeForce RTX 5060 Ti 16 GB16 GB~169 tok/slow $780, avg $1,146 (new) source, 2026-10-16
NVIDIA GeForce RTX 4080 SUPER16 GB~279 tok/scheck price
NVIDIA GeForce RTX 408016 GB~271 tok/scheck price
NVIDIA GeForce RTX 4070 Ti SUPER16 GB~254 tok/scheck price
NVIDIA GeForce RTX 4060 Ti 16 GB16 GB~109 tok/scheck price
AMD Radeon RX 9070 XT16 GB~244 tok/slow $810, avg $1,194 (new) source, 2026-10-16
AMD Radeon RX 907016 GB~244 tok/slow $570, avg $1,052 (new) source, 2026-10-16
AMD Radeon RX 9060 XT 16 GB16 GB~122 tok/slow $620, avg $812 (new) source, 2026-10-16
AMD Radeon RX 7800 XT16 GB~236 tok/slow $599, avg $834 (new) source, 2026-10-16
AMD Radeon RX 7600 XT16 GB~109 tok/scheck price
AMD Radeon RX 7900 XT20 GB~303 tok/scheck price
NVIDIA GeForce RTX 409024 GB~382 tok/slow $3,200, avg $4,663 (used) source, 2026-10-16
NVIDIA GeForce RTX 3090 Ti24 GB~382 tok/scheck price
NVIDIA GeForce RTX 309024 GB~354 tok/slow $1,889, avg $2,404 (used) source, 2026-10-16
AMD Radeon RX 7900 XTX24 GB~363 tok/slow $1,350, avg $1,552 (new) source, 2026-10-16
NVIDIA Tesla P4024 GB~131 tok/scheck price
NVIDIA GeForce RTX 509032 GB~677 tok/slow $6,500, avg $8,757 (new) source, 2026-10-16
AMD Radeon AI PRO R970032 GB~244 tok/scheck price
Apple Mac Studio M5 Max (32-core GPU, 36 GB)36 GB~174 tok/slow $2,499 (new) source, 2026-10-16
NVIDIA RTX A600048 GB~291 tok/slow $6,170 (new) source, 2026-10-16
NVIDIA RTX 6000 Ada Generation48 GB~363 tok/scheck price
Apple Mac Studio M5 Max (40-core GPU, 64 GB)64 GB~232 tok/scheck price
Apple Mac Studio M5 Ultra (64-core GPU, 96 GB)96 GB~454 tok/scheck price
Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB192 GB~103 tok/slow $6,799 (new) source, 2026-10-16

Speed calculator estimate

Estimate = 0.76 × bandwidth ÷ GB read per token (Q4 file size, scaled by active/total parameters for MoE). Fit = file size + 1.5 GB headroom ≤ total VRAM. Multi-GPU: VRAM adds, speed uses one card's bandwidth (layer split). Short context only. How this works

Related